Research map / 3D vision / Human pose estimation
First-Person Hand Action Benchmark With RGB-D Videos and 3D Hand Pose Annotations
Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, Tae-Kyun Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hand pose features to depth descriptors (ablation) | HOG²-depth · First-person hand action dataset | accuracy (%): 59.83 -> 66.78 | Improved |
| Combine color, depth, and hand pose (ablation) | JOULE · First-person hand action dataset | accuracy (%): 74.60 -> 78.78 | Improved |
| Use two LSTM layers instead of one (ablation) | LSTM action recognition · First-person hand action dataset | accuracy (%): 78.73 -> 80.14 | Improved |
| Increase the training share from 1:3 to 1:1 (ablation) | LSTM action recognition · First-person hand action dataset | accuracy (%): 58.75 -> 78.73 | Improved |
| Evaluate on a held-out person rather than a sequence-level split (ablation) | LSTM action recognition · First-person hand action dataset | accuracy (%): 84.82 -> 62.06 | Hurt |
| Combine hand pose with object pose (ablation) | LSTM action recognition · Object-pose subset | accuracy (%): 87.45 -> 91.97 | Improved |
| Use object pose instead of hand pose (ablation) | LSTM action recognition · Object-pose subset | accuracy (%): 87.45 -> 74.45 | Hurt |
| Train the hand pose estimator with hand-object images (ablation) | Hand pose estimation and LSTM action recognition · First-person hand action dataset, action split | action accuracy (%): 29.63 -> 72.06 | Improved |
| Train with objects represented in both splits rather than hold out objects (ablation) | Hand pose estimator · First-person hand action dataset | pose error (mm): 19.84 -> 11.25 | Improved |
| Use estimated hand poses instead of ground-truth magnetic poses (ablation) | LSTM action recognition · First-person hand action dataset | action accuracy (%): 78.73 -> 72.06 | Hurt |
| Use estimated poses with Lie Group instead of ground-truth poses (ablation) | Lie Group action recognition · First-person hand action dataset | accuracy (%): 82.69 -> 69.22 | Hurt |
| Use estimated poses with Gram Matrix instead of ground-truth poses (ablation) | Gram Matrix action recognition · First-person hand action dataset | accuracy (%): 85.39 -> 32.22 | Hurt |
Closest papers by meaning
- InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image (ECCV 2020 · similarity 0.92)
- On the Utility of 3D Hand Poses for Action Recognition (ECCV 2024 · similarity 0.92)
- ComPose: When to Trust Hands for Object Pose Tracking (NeurIPS 2026 · similarity 0.90)
- How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions (CVPR 2025 · similarity 0.90)
- HOLD: Category-agnostic 3D Reconstruction of Interacting Hands and Objects from Video (CVPR 2024 · similarity 0.89)
- HOT3D: Hand and Object Tracking in 3D from Egocentric Multi-View Videos (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
